Ref: https://learn.cantrill.io/courses/1820301/lectures/41301607

• Summary
• Contents:

EC2 Placement Groups - Overview

• 💡 The physical placement of an EC2 instance is normally determined by AWS
  • usually in the host that makes most sense, within the AZ that the instance is launched in
• 🔧 Placement groups influence how instances inside a group are placed physically in EC2
  • They can ensure that instances are physically together or separated
• 3 types of placement groups:
  1. Cluster: pack instances physically close together
  2. Spread: keep instances separated (instances use different underlying HW)
  3. Partition: instances grouped physically in partitions, but partitions are spread apart

Cluster Placement Groups

Diagram: https://github.com/acantril/aws-sa-associate-saac03/blob/main/1100-ADVANCED_EC2/00_LEARNINGAIDS/PlacementGroups-1.png

• 🔧 Concept: Instances tightly packed together, usually to achieve high performance
  • Instances use the same infrastructure rack, and often also the same host
  • ❗ Not all instance types supported
• ❗ Group is launched in a SINGLE AZ → can't span AZs
  • 💡 Being in different AZs would degrade performance
  • AZ locked when first instance is launched
  • Group can span VPC peers… but it significantly impacts performance in a negative way!
• Using same instance type not mandatory, but provides best results
• ‼️ Launching all instances at the same time is highly recommended (best practice)
  • Ensures AWS allocates necessary capacity for all that you require
  • 💡 Certain capacity provided for first instances, adding more later can be problematic
• 👍 Main benefit: instances get fast & direct BW to all other instances inside the same group
  • Highest throughput, lowest latency, maximum pps (packets/s) possible in AWS
    • ‼️ Use instances with high performance NWing + enhanced NWing on all instances → maximizes the benefit of this placement group
  • Single data stream transfer rate up to 10Gbps (vs normal 5Gbps)
    • 💡 some instances have faster rates, but transfer still limited by single stream rate
• 👎 Main drawback: little to no resilience
  • if HW fails, it can take down all instances
• Use cases: systems that require high performance, fast speeds & low latency
  • e.g. High Performance Compute (HPC) or scientific analysis

Spread Placement Groups